US20030208289A1 - Method of recognition of human motion, vector sequences and speech - Google Patents

Abstract

A method for recognition of an input human motion as being the most similar to one model human motion out of a collection of stored model human motions. In the preferred method, both the input and the model human motions are represented by vector sequences that are derived from samples of angular poses of body parts. The input and model motions are sampled at substantially different rates. A special optimization algorithm that employs sequencing constraints and dynamic programming, is used for finding the optimal input-model matching scores. When only partial body pose information is available, candidate matching vector pairs for the optimization are found by indexing into a set of hash tables, where each table pertains to a sub-set of body parts. The invention also includes methods for recognition of vector sequences and for speech recognition.

Description

RELATED APPLICATIONS
• This application claims the benefit of U.S. Provisional Application No. 60/378,316 filed May 6, 2002. And claims the benefit of U.S. Provisional Application No. 60/381,002 filed on May 15, 2002. The entire teachings of these are incorporated herein by reference.[0001]
STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH AND DEVELOPMENT
• [0002] Parts of research related to the invention were sponsored by the National Science Foundation (NSF). According to the NSF patent regulations (45 CFR 650.9), the NSF gave me a waiver and agreed that I shall retain the rights to the invention and file a patent application. My case number is CV78 NSF Grant No. 02-454.
BACKGROUND OF THE INVENTION
• A. Field of the Invention[0003]
• The invention is a novel method for recognition of human motions. The invention also applies this method for vector sequence recognition and for speech recognition. Suggested classification number for the invention is: “73/865.4—Measuring and Testing/Analyzing Bodily Movement”.[0004]
• B. Description of Prior Art[0005]
• Recognition of human motion and especially recognizing detailed human activities, is a relatively new research area with very few published works. A paper by Ben-Arie, J., Wang, Z., Pandit, P. and Rajaram, S., “Human Activity Recognition Using Multidimensional Indexing,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), Vol. 24, No. 8, pp. 1091-1105, August 2002 is among these few. An additional excellent reference which conducts an intensive survey on the different methodologies for visual analysis of human movement is in D. M. Gavrila. “The visual analysis of human movement: A survey,” Computer Vision and Image Understanding, 73(1):82-98, 1999. Gavrila groups them into 2D approaches with or without explicit shape models and 3D approaches. The 2D approach without explicit shape models is based on describing human movement in terms of simple low-level 2D features instead of recovering the pose. The second approach, which is a view-based approach uses explicit shape models to segment, track and label body parts. The third approach attempts to recover the 3D poses over time. More recently, there has been a survey by T. B. Moeslund and E. Granum “A survey of computer vision-based human motion capture,” Computer Vision and Image Understanding, 81(3):231-268, March 2001, which describes various computer vision-based human motion capture. They elaborate about the various categories of human motion capture namely Initialization, Tracking, Pose Estimation and Recognition. Human motion recognition is classified into static and dynamic recognition. Static recognition is based on using spatial data, one frame at a time and dynamic recognition uses the temporal characteristics of the action. Our method of Recognition Indexing & Sequencing (RISq) is based on a novel approach that differs from all the methods surveyed above.[0006]
• Unlike our method, which can classify many different activities, past works focused on recognition of only few activity classes. H. Fujiyoshi and Alan J. Lipton in “Real-time human motion analysis by image skeletonization,” Proc. of the Workshop on Application of Computer Vision, October 1998, use skeletonization to extract internal human motion features and to classify human motion into “running” or “walking” based on the frequency analysis of the motion features.[0007]
• M-H. Yang and N. Ahuja. in “Recognizing hand gesture using motion trajectories,” IEEE Conference on Computer Vision and Pattern Recognition, pages 466-472, June 1999, apply Time-Delay Neural Network (TDNN) to hand gesture recognition and achieve quite high recognition rate a method akin to Dynamic Time Warping (DTW). DTW was also used by works such as: Sakoe H., Chiba S., in “Dynamic Programming Optimization for Spoken Word Recognition”, IEEE Transactions on Acoustics, Speech, and Signal Processing, 34(1):4349, 1978.J. Schlenzig, E. Hunter, and R. Jain in “Vision based hand gesture interpretation using recursive estimation,” Proceedings of the 28th Asilmoar Conference on Signals, Systems and Computers, 1994, use Hidden Markov Model (HMM) and a rotation-invariant imaging representation to recognize visual gestures such as “hello” and “good-bye”. HMMs are used by J. Yamato, J. Ohya, and K. Ishii in “Recognizing human action in time-sequential images using hidden markov model,” in Proceedings Conference on Computer Vision and Pattern Recognition, pages 379-385, June 1992, for recognizing human action in time sequential images. HMMs were also utilized by Starner and Pentland to recognize American Sign Languages (ASL). Darrell and Pentland applied dynamic time warping to model correlation for recognizing hand gestures from video. R. Polana and R. Nelson in “Recognizing activities” Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition, pages 815-818, 1994, use template matching techniques to recognize human activity.[0008]
• Motion Energy Images are used by A. F. Bobick and J. W. Davis in “An appearance based representation of action” in Proc. of Thirteenth International Conference on Pattern Recognition, August 1996, for recognition. I. Haritaoglu, D. Harwood, and L. S. Davis in “w4: Real-time surveillance of people and their activities⇄ in IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):809-830, August 2000, implemented a system for human tracking and activity recognition, in which the activity recognition part is mainly based on analysis of the projected histograms of detected human silhouettes. This system classifies human poses in each frame into one of four main poses (standing, sitting, crawling/bending, lying) and one of three view-based appearances (front/back, left-side and right side) and activities are monitored by checking the pose changes over time.[0009]
• In another work, Y. A. Ivanov and A. F. Bobick in “Recognition of visual activities and interactions by stochastic parsing” in IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):852-871, August 2000, recognize generic activities using HMM and stochastic parsing. These activities are first detected as a stream of low level action primitives represented using HMM and then are recognized by parsing the stream of primitive representations using a context-free grammar. A. F. Bobick and J. W. Davis in “The recognition of human movement using temporal templates” in IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(3), March 2001 recognized human activity by matching temporal templates against stored instances of views of known actions.[0010]
• More recently, Galata et al. [9] A. Galata, N. Johnson, and D. Hogg in “Learning variable-length Markov models of behavior” in Computer Vision and Image Understanding, 81(3):398-413, March 2001, use Variable Length Markov Models (VLMM) for modeling human behavior. They use VLMMs because of their more powerful encoding of temporal dependencies. Our indexing based recognition RISq approach differs from all the above-mentioned works since it determines the best matching activity in a single indexing operation, is invariant to activity's speed and requires only few very sparse samples of the activity for complete recognition.[0011]
• String matching techniques that allow recognition of input strings as similar to stored collection of model strings stored in a database, are also described in several patents such as U.S. Pat. No. 5,577,249 or in U.S. patent application Ser. No. 20,020,181,711. However, they significantly differ in their method from our method. U.S. Pat. No. 5,577,249 employs random partitioning for the recognition and U.S. patent application Ser. No. 20,020,181,711 finds similarity by generating K-means cluster signatures.[0012]
• Other patents that relate to human motion recognition are mostly adapted to recognition of simple actions such as hand gestures as an input to a computer etc. and not for recognition of human motion such as articulated human activity. They also have recognition methods that substantially differ from our method. Such are U.S. Pat. Nos. 6,256,033, 5,581,276, 5,714,698 and 6,222,465. U.S. patents that relate to indexing and speech recognition such as: U.S. Pat. Nos. 5,386,492, 5,502,774, 5,621,809, 6,292,779, 6,542,869, 6,371,711, and U.S. patent application Ser. Nos. 20,020,052,742 or 20,020,062,302 apply methods that are different from our invention.[0013]
• The prior art fails to provide satisfactory solutions to a cardinal problem in the recognition of human motion. Human motion has noticeable variations in speed that occur all the time. The speed varies between person to the next and even during performance of the same person. All the prior methods are based on equally fast rate of sampling of both input and model motions. Prior recognition algorithms take into account the temporal properties of the motion since they rely on temporal correlation as the basic principle of matching, whether the algorithm is continuous or discrete. The only prior method that tried to cope with speed variations was Dynamic time warping (DTW) this is a dynamic programming technique used to create nonlinear warping function between the input time axis and the model time axis. DTW is substantially different from our method requires a lot of computations that not always converge and also is still quite sensitive to large variations of speed. In contrast, our invention proposes a method that eliminates entirely the time factor from the recognition and replaces it with sequencing. Another innovation of our approach, i.e. the substantially different rate of sampling, enables to recognize human motions just from very few samples. This result was verified by us experimentally. Similar advantages are expected form the application of the method towards speech recognition.[0014]
SUMMARY OF THE INVENTION
• The present invention is a novel method that enables recognition of an input human motion as being the most similar to one model human motion out of a collection of model human motions that are stored in a database. The invention also includes a general method for recognition of vector sequences and a method for recognition of speech.[0015]
• For the recognition of human motion, both the input and the model human motions are represented by vector sequences that are derived from samples of recorded and measured trajectories of human body parts while performing the motions. In the preferred embodiment of the invention the trajectories are of the angular poses of human body parts. Since the skeletal structure of the human body consists of bones connected to one another by flexible joints, the angular poses of human body parts uniquely determine the entire posture of the human body. Hence, a sequence of samples of the angular poses of body parts and their derivatives, provide all the relevant information about the motion being performed.[0016]
• In the preferred embodiment of the invention with regards to recognition of human motion, we first measure and sample the angular poses of the body parts in performing model motions. Next, we represent them by model vectors that include also the derivatives of each pose. After that, we separate the model vectors into sub-vectors that correspond to the sub-sets of body parts and store the model sub-vectors of each subset in a separate hash table. Then, we index separately into each hash table using input motion sub-vectors that were acquired from the input motion, in the same manner as the model sub-vectors, and retrieve the model sub-vectors that are similar to the input sub-vectors. The next step, is to merge the model vectors retrieved and to find the best matching of the input vector sequence with each of the models. The optimization algorithm used is based on dynamic programming and also employs a special procedure that includes only matching sequences that fulfill the sequencing requirement. The model with the highest matching score with the input sequence is recognized as the motion class most similar to the input. The methods for vector sequence recognition and for speech recognition are similar, but do not divide the vectors into sub-vectors.[0017]
• One of the main innovative features of the invention is the use of sampling rates which are substantially different for the input motion and for the model motions. In one option, we propose to sample the input motion with slow-sparse rate and the model motion with fast rate. In the second option we propose to switch the rate, i.e. a fast sampling rate for the input and slow-sparse rate for the models. The reasons behind the substantial difference in sampling rates are quite important and provide significant advantages.[0018]
• Firstly, it is not required to sample both the input and the models with fast rate of sampling—as all the other motion recognition methods do. In fact, when one kind of motion is sampled sparsely, the recognition of human motion is still quite robust, and with probability for errors, which is very low.[0019]
• Secondly, matching a sequence of slow-sparse samples to a dense sequence of samples (obtained by fast sampling rate) enables to recognize input motions that are in the same class as one of the models, but significantly differ in their speed. This is because in our approach, the time periods between the sparse samples are of no consequence, since each sample of the sparse sequence is matched to a sample of the dense (fast) sequence without relevance to their timing. The speed can even vary during the motion and our method will still be able to recognize the motion. This is important since the same human motions in performing the same actions, naturally vary. Other recognition methods did not eliminate the time factor from the motion, as we do in this invention, and therefore have severe problems when the input speed does not match the model speed.[0020]
• Thirdly, sparse sampling requires to match only very few sparse samples for robust recognition. Practically, only 4-5 sparse pose samples are required for highly accurate recognition of human motions, as we found in experiments elaborated in our paper: Ben-Arie, J., Wang, Z., Pandit, P. and Rajaram, S., “Human Activity Recognition Using Multidimensional Indexing,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), Vol. 24, No. 8, pp. 1091-1105, August 2002. This results in noticeable savings in the computations required. Both the first and the second options offer similar efficiency in reducing the computational load in the recognition process.[0021]
• Another innovative feature of the invention is the requirement for sequencing in matching two vector sequences. The sequencing requirement do not allow to match two vector sequences which are not in the same order. Gaps are allowed but not reversals in ordering of matching vector pairs. It means that the imaginary arcs that connect each input vector to its matching model vector in the matching vector pair, never cross another imaginary arc that belong to another vector pair. This requirement is of great consequence since it guarantees that matching vector pairs that contribute to the matching scores of sequences matched must have the same sequence. It means that matching motions must have the same sequence of poses, but not necessarily the same speed. Actually, the sequencing requirement along with the large variation in sampling rates, were introduced in order to eliminate entirely the time factor in matching vector sequences. As mentioned in paragraph [0017], this enables to recognize human motions, vector sequences and speech that vary in their speed relative to their stored models.[0022]
• An additional novel feature of the invention is the requirement that vectors included in the sparse sequence, are matched only to one vector in the fast sequence. This requirement severely constrains the number of possible vector pair sequences in the optimization.[0023]
• In addition to previous features, another unique and novel feature of the invention is the introduction of separate hash tables and separate indexing for each sub-set of body parts. This feature allows to recognize human motions even if poses of several body parts are not known due to occlusion or missing measurement information. This feature is quite important in practice, since missing pose information occurs frequently.[0024]
• As far as we have searched, all the 4 features of our invention, mentioned in paragraphs [0015], [0019], [0020], [0021] above are innovative and unique.[0025]
BRIEF DESCRIPTION OF THE DRAWINGS
• FIG. 1 is a schematic block diagram describing the preferred steps suggested in order to create 5 hash tables that store a collection model sub-vectors. These 5 sub-vectors represent the poses of the 4 limbs and the torso of humans in performing a corresponding collection of model human motions.[0026]
• FIG. 2A is a schematic block diagram illustrating the first 5 steps suggested as a preferred method for recognition of an input human motion.[0027]
• FIG. 2B is a schematic block diagram illustrating the next 5 steps, following the steps in FIG. 2A, that are suggested as a preferred method for recognition of an input human motion.[0028]
• FIG. 2C is a schematic block diagram illustrating the subsequent 3 steps, following the steps in FIG. 2B that are suggested as a preferred method for recognition of an input human motion.[0029]
• FIG. 3 is the schematic drawing of the data structure used in the process of sequence matching. This data structure consists of 5 variables stored in a cell. The cell is a component of array A[0030]_j.
• FIG. 4 is the flow chart describing the construction of the array set {A[0031]_j}. Each array A_jin the set corresponds to another model M_j.
• FIGS. 5A and 5B illustrate the flow chart of a preferred optimization algorithm for finding of the best matching score among all the possible valid sequences of matching vector pairs (valid pair sequences), constructed from vector pairs composed of a given input vector sequence T[0032]_kand one model vector sequence M_j. The algorithm shown illustrates a preferred embodiment of the invention. The algorithm also guarantees the pair sequence selected, is valid (i.e. fulfills the sequencing requirement).
DETAILED DESCRIPTION OF THE INVENTION
• Mention speech and the 2 options at the end, explain signal transforms for speech.[0033]
• This invention is mostly focused on methods for recognition of human motion. However, similar methods are also proposed for other applications that include speech recognition and recognition of vector sequences per se.[0034]
• Hence, the major objective of the invention is to recognize inputs of human motion and to classify each one of these inputs as most similar to one member of a collection of stored model human motions. The manner of acquisition of the motions does not perturb the principles of the invention. Human motion can be measured and tracked by a large variety of systems and sensors. Depending whether these systems are attached to the subject's body or is installed in the subject's environment, these systems can be classified as internal or external. Examples of external measurement systems are 2D or 3D video, optical tracking of markers such as infra-red or LEDs (light emitting diodes), gravitational or inertial sensors, electromagnetic sensors and others. Examples of internal systems are data-gloves or overall suits equipped with devices such as piezo-resistive flex sensors. Such devices can measure the relative orientation between various body parts. There is a large variety of other devices that are applied for relative orientation measurement. These include fiber optic or light-tube flexes, potentiometers, shaft encoders, short range mutual inductance devices, Hall effect devices, electro-myographic devices (instruments that measure muscular contractions) and others. Most of these devices are quite small and do not require external source, thus provide almost unlimited space in which the person can move and perform. In our preferred embodiment we suggest to acquire the motion by tracking body parts in video.[0035]
• To recognize human motion one ought to acquire detailed information that pertains to the movements of all the body parts involved in such motion. For such a purpose, the human body is usually represented by a skeletal model of all the bones, which are connected by flexible joints. Thus, motion of body parts such as arms, legs or fingers can be described as motion of the corresponding skeletal part. Since every bone is connected to another bone of the skeleton via a flexible joint, the skeletal structure can be represented by a tree graph. Where the tree's nodes represent bones and the tree's arcs correspond to joints. The root of the tree is usually defined as the sacrum (or sacroiliac joint) and all the other bones are siblings of the root. The pose in 3D space of any bone can be decomposed into 3D position and 3D orientation. Correspondingly, movement of any bone can be described as 3D translation plus 3D rotation. But, since all the bones are connected to one another by flexible joints, their translations are not independent and are a function of the predecessor's translation of each bone in the tree structure. Hence, the only independent 3D translation is of the root's. All the other body's parts independent poses and motions are determined entirely by their 3D angular orientations and their 3D angular velocities. Such a representation is employed by the widely used Virtual Reality Modeling Language (VRML). The VRML is mostly used for computer animation of humanoids and avatars and in video compression.[0036]
• The rudimentary representation of human motion is by the trajectories of various body parts that participate in the motion. These trajectories describe the 3D poses of the body parts as functions of time. We find that major human activities such as walking, running, jumping etc. can be characterized and later robustly recognized by the trajectories of 9 major body parts. These include the upper arms, the forearms, the thighs, the calves, and the torso. Evidently, there are many other human motions that can be accurately represented by 3D trajectories of other body parts such as the hands, fingers etc. As elaborated above, the 3D poses and velocities of body parts that participate in human motion, can be accurately described by the angles of their skeletal joints and their derivatives.[0037]
• FIG. 1 describes the suggested steps of the preferred embodiment of our method of acquiring a collection of model human motions. The various models of human motion stored in this collection represent corresponding classes of human motion. These models are later tentatively matched with an input motion of unknown class. The matching quality is expressed by matching scores, which reflect the similarity between each model and the input motion. The input motion is recognized as most similar to the model with the highest matching score.[0038]
• In our preferred embodiment, we suggest to use an inverse function of the distance between the vectors that represent the input motion i.e. t[0039]_nkand the corresponding model vector m_rj. Such an inverse function, reflects the quality of matching, or in other words, the similarity between t_nkand m_rj. In our preferred embodiment, we employ an inverse function such as 1/[a+d(t_nk, m_rj)], where d(t_nk, m_rj) is the Euclidean distance between t_nkand m_rjand a is a positive constant.
• The motion models are acquired by asking human subjects to perform various activities in a controlled environment. This step is described by[0040]block101. Next, tracking techniques are used to obtain the trajectories of the 3D poses of the major body parts that participate in the motions modeled. This step is described byblock102. The next step, described inblock103, is to sample the trajectories of all the body parts tracked. The sampling rate for the model motion can be significantly different from the sampling rate used for the input motion.
• In this invention, we propose two options for the sampling rates. The first option is that the input motion is sampled at a slow-sparse sampling rate and the collection of model motions are sampled at a fast rate. In the second option, the sampling rates are switched. The reasons behind the introduction of different sampling rates in this invention are quite cardinal, and actually provide important advantages. Firstly, it is not necessary to sample both the input and the models with fast rate of sampling—as all the other motion recognition methods do. In fact, recognition of human motion is quite robust, and with very low probability for errors, even when one kind of motion is sampled sparsely. Secondly, matching a sequence of sparse samples to a dense sequence of samples (obtained by fast sampling rate) enables to recognize input motions that are in the same class as one of the models, but significantly differ in their speed. The speed can even vary during the motion and our method will still be able to recognize the motion. This is important since the same human motions in performing the same actions, naturally vary in their speeds from one person to the other. Thirdly, sparse sampling requires to match only few sparse samples for robust recognition. This results in noticeable savings in the computations required. Additional details on the different sampling rates are provided in paragraphs [0056][0057][0058].[0041]
• The next step in the construction of the hash tables for the collection of model motions, is described in[0042]block104. In this step, the sampled poses of human body parts are converted into joint angles and joint angular derivatives. Computing transformations from sampled poses to angles is necessary if the sampled poses are not originally measured in angles. For example, if the poses of body parts are measured by their 3D locations, the joint angles can be computed from these locations if the anthropomorphic proportions or measures of the human subject performing the motions are known.
• The angular derivatives are usually extracted from the differences of two or more successive samples. The order of the derivatives determines the minimal number of successive samples that are needed. For first derivatives two successive samples are the minimum and the derivative is computed from the difference between the samples. For second derivatives, 3 samples are the minimum, and so on. The second derivative d″(t) is proportional to d″(t[0043]₂)=d(t₁)−2d(t₂)+d(t₃) where d(t₁) is the first sample, d(t₂) is the second and d(t₃) is the third. However, we do not see practical use for derivative orders higher than 2. Computing derivatives for slow-sparse sampling requires to take two or more successive samples at fast rate at any point in time that the slow-sparse sample is taken. Thus slow-sparse sampling actually results in “bursts” of two or three fast samples that are separated by relatively long time periods.
• In the preferred embodiment described in[0044]block104, the poses of the 9 major body parts and their first derivatives, are transformed into vectors of 18 dimensions. For the model motions, these vectors are denoted by m_rjand the model vector sequence that represents the entire model motion is denoted by M_jwhere subscript j is the model number in the collection, and subscript r is the location of m_rjwithin the sequence.
• In the preferred embodiment that is described by[0045]block105, it is suggested to separate the model vectors into 5 model sub-vectors. Four sub vectors have 4 dimensions that represent the angular poses and their derivatives of the 4 limbs. Four dimensions are required since the each of the limbs has two major body parts. The legs are divided into thighs and calves and the arms are divided into upper arms and forearms. The sub-vector that pertains to the torso is just two dimensional. The reason behind this separation is that it enables to vote separately on each body part and if one or more body parts are occluded, the matching scores aggregated from the other body parts can still amount to robust recognition. In the preferred embodiment of other applications such as speech or vector sequence recognition, such a separation is not necessary.
• The sub-vectors sequences resulting from[0046]105 are stored in 5 separate hash tables. The creation of these hash tables is described inblock110. These hash tables have bins that each is addressed by a different address vector. The address vectors for the 4 dimensional hash tables are 4 dimensional and for the 2 dimensional hash table are 2 dimensional. The values of the address vectors for each hash table can be obtained by a Vector Quantization (VQ) algorithm. In the VQ algorithm, all the sub-vectors that belong to the table from all the model sequences are used to find a small, predetermined number of quantized vectors that represent the whole sub-vector collection with minimal mean square error. These quantized vectors are then used as address vectors for the bins. Whenever, the table is addressed, the sub-vectors stored in the addressed bin are retrieved. All the retrieved vectors are similar to the addressing vector in the sense that they have relatively low distance from the addressing vector. These retrieved model sub-vectors are merged to a single model vector for each matching (t_nk, m_rj) and their vector matching scores are evaluated after merging. The reason behind this structure is to enable recognition when some of the body parts are missing. In such a case, only the body parts with known sub-vectors are combined.
• Since both input and model motions are represented by vector sequences, this task is equivalent to finding the best matching between a given input vector sequence and one member of a collection of model vector sequences. The preferred embodiment of the invention involves a recognition method that employs indexing into hash tables followed by an optimization-sequencing algorithm that finds the valid-sequence of matching vector pairs with the highest total sum of vector matching scores. The initial preparation of the hash tables used in the algorithm is illustrated in FIG. 1, which is described above. FIG. 2A, FIG. 2B, and FIG. 2C outline the various steps involved in the recognition process. The preferred approach to solve the optimization-sequencing problem is described with the help of drawings FIG. 3, FIG. 4 and FIG. 5A and FIG. 5B.[0047]
• As mentioned above in paragraph [??], FIG. 2A, FIG. 2B and FIG. 2C describe the general steps required in the recognition process. The actions that are described in[0048]steps201,202,203,204 and205 correspond exactly to the steps described insteps101,102,103,104 and105 in FIG. 1, except that now they are implemented on the input motion instead. The model vectors in FIG. 1 are replaced by the input motion vectors t_nkand the input vector sequence is denoted here by T_kwhere the subscript k denotes the serial number of the input vector sequence and the subscript n signifies the location of t_nkwithin T_k.
• In[0049]block210 the subscript n is initialized to point the first input vector t_1kwhose sub-vectors are used inblock211 for indexing.Block211 describes the indexing of the 5 input sub-vectors into the 5 hash tables that takes place as an initial step in the recognition. Each such indexing yields groups of 5 model sub-vectors that are similar to the 5 input sub-vectors used for the indexing. Each of the groups of 5 model sub-vectors retrieved, are merged into a complete model vector and coupled with the input vector that was used for the indexing, to create a vector pair (t_nk, m_rj) inblock212. Inblock213, the matching scores of these vector pairs are computed and the pairs are removed if the pair matching score is smaller than a predetermined threshold.
• In[0050]block214, k is incremented by 1 and if k is not greater than p the algorithm returns to block211. And if k is greater than p the algorithm moves to block215. Inblock215 the optimal matching score between the input vector sequence T_kand the model vector sequence M_jis found using the optimization algorithm that is described in detail in FIGS. 3, 4,5A and5B. The optimization algorithm also guarantees that the pair sequence found is also valid, i.e. it fulfills the sequencing requirement, detailed in paragraph [**bb]. Inblock216 the subscript j is incremented by 1 and if j is not greater than z, which is the total number of models in the collection of model vector sequences {M_j} the algorithm returns to block211. And if j is greater than z the algorithm moves onwards to block216. Inblock220 the optimal matching scores that pertain to all the z models in the collection of model motions are compared. The model with the highest matching score is selected as the one that is most similar to the input motion i.e. the input motion is recognized as belonging to the class of model motions represented by the winning model.
• In order to find the highest total sum of vector matching scores between the input vector sequence and a model vector sequence, we use in the optimization-sequencing algorithm a 5 variable data set that is attached to each input-model matching vector pair. These vector pairs, which are created from any combination of input and model vectors that have a matching score above some predetermined threshold. These matching vector pairs are arranged in a sequence represented by the array A[0051]_j. FIG. 3 describes the data set henceforth called the cell. The cell is the basic building block of array A_j, which is an array of cells and will be described in the subsequent section??. Each cell corresponds to a matching input-model pair and includes the values of 5 variables that represent the following information:
• (1)[0052]301—the first variable is Previous_cell, it is a pointer to the location of the previous cell in the array A_j.
• (2)[0053]302—the second variable is O_nkrj, which is the partial sequence matching score indicating the highest sum of vector matching scores of a partial vector pair sequence that ends with the cell related matching vector pair, i.e. the pair of input vector matching with model vector that are related to the cell.
• (3)[0054]303—the third variable is W_nkj, which stores the vector pair's matching score, i.e. the matching score of the input motion vector and model motion vector that are related to the cell,
• (4)[0055]304—the fourth variable is R, which stores the location in the sequence M_jof the model vector related to the cell.
• (5)[0056]305—the fifth variable is N, which stores the location of the input vector within the input motion sequence T_k.
• The array A[0057]_jas mentioned earlier is made up of cells and contains one cell for each vector matching pair composed of the input vector sequence and the model vector sequence in consideration.
• The construction of the array A[0058]_jis illustrated in the flow chart FIG. 4. In order to describe this construction process and subsequent matching process, we assume that the input vector sequence T_kcontains the sequence of vectors (t_1k, t_2k. . . t_nk. . . t_pk), which is being matched with the model vector sequence M_jcontaining the sequence of vectors (m_1j, m_2j. . . m_rj. . . m_qj). The variable j holds the information of the current model motion in consideration. It is initialized to 1 as depicted in401. The array A_jis initialized402 by adding to it a null cell that depicts the null matching of t_0kwith m_0j, which are two null vectors. All the variables of the null cell are set to zero. Cells are appended to the array A_jfor each matching pair of t_nkwith m_rjwith matching score above a certain threshold as depicted in403. This process is repeated for all input vectors t_nkin the sequence T_kand stops when all the vectors in the input vector sequence has found all the matching model vectors in all the z models in the collection—as illustrated in404 and405. At theend410 of this loop, the array A_jholds all the required information for obtaining the highest sum of matching scores between a given input vectors sequence and the model vector sequence M_j. This construction procedure for A_jhas to be repeated for each model vector sequence M_jthat belongs to the collection of model vector sequences {M_j} to form the set {A_j}.
• As mentioned above, the objective of our optimization-sequencing algorithm is to find the highest sum of matching scores between the input vector sequence T[0059]_kand model vectors sequence M_j. The algorithm actually finds a sequence of matching input model vector pairs which has the highest total sum of individual pair matching scores. The optimal sequence of matching vector pairs has also to fulfill a sequencing requirement that is defined next, in order to be considered as a valid vector pair sequence. The sequencing requirement has two options depending on which of the matching vector sequences is acquired at a slow-sparse sampling rate and which at a fast sampling rate.
• The first option is that the input vector sequence T[0060]_kis acquired at a slow-sparse rate of sampling and all the models in the collection {M_j} are acquired at a fast rate of sampling. By the term “slow-sparse rate of sampling” we mean that the input human motion is sampled at timing instances that are separated by relatively long periods of time compared to the time periods that separate the “fast rate of sampling”. It does not mean that the slow-sparse speed has necessarily a uniform slow rate. Sometimes it is advantageous to sample human motion at particular instances, when the pose of body parts is at extreme position or when the pose is indicative of the kind of motion being performed. A typical range for the time periods between samples is 0.4-1.2 seconds. In our experiments with human motion recognition we found that such sample spacing range was quite useful. The fast sampling rate in our experiments was uniform at about 30 samples per second. One of the major advantages of the method is that only few sampled human poses are required for correct recognition of human motion with very high probability. This property is proved in our paper Ben-Arie, J., Wang, Z., Pandit, P. and Rajaram, S., “Human Activity Recognition Using Multidimensional Indexing,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), Vol. 24, No. 8, pp. 1091-1105, August 2002.
• Assuming in the first option, that the input motion is sampled at a slow-sparse rate and the collection of model motions are sampled at fast rate, the sequencing requirement demands that any possible permutation of two matching vector pairs . . . (t[0061]_nk, m_rj) . . . (t_ck, m_ej) . . . included in a valid vector-pair sequence, has to fulfill the following 3 conditions: (I) c is not equal to n; (II) if c>n, then r≦e; (III) if c<n then e≦r. Any total matching score ought to include only pair matching scores that pertain to matching vector pairs that are included in a valid vector-pair sequence. The sequencing requirement actually guarantees that the imaginary arcs that connect each input vector to its matching model vector of the pair, never cross one another and that each input vector t_nk, which is included in a valid vector-pair sequence, is matched only to one model vector m_rj.
• Assuming in the second option, that the collection of model motions are sampled a at slow-sparse rate and the input motion is sampled at a fast rate, the sequencing requirement demands that any possible permutation of two matching vector pairs . . . (t[0062]_nk, m_rj) . . . (t_ck, m_ej) . . . included in a valid vector-pair sequence, has to fulfill the following 3 conditions: (I) r is not equal to e; (II) if e>r, then n≦c; (III) if e<r then c≦n. Any total matching score ought to include only pair matching scores that pertain to matching vector pairs that are included in a valid vector-pair sequence. The sequencing requirement actually guarantees that the imaginary arcs that connect each input vector to its matching model vector of the pair, never cross one another and that each model vector m_rj, which is included in a valid vector-pair sequence, is matched only to one input vector t_nk.
• Both the first and the second options offer similar efficiency in reducing the computational load in the recognition process. In addition, the use of sparse-slow samples that are matched to fast samples conserving only the sequencing requirement, allows to match two vector sequences of the same motion but with entirely different speeds, or even with significantly varying speeds during the motion. Such a flexibility is quite important especially in recognition of human motion or speech that may vary in their speed from one person to the other or even with the same person.[0063]
• The optimization algorithm is described with the help of flow charts in FIG. 5A and FIG. 5B. The optimization principle employed here, is based on Dynamic Programming (DP). Finding the optimal matching between an input vector sequence and a model vector sequence, is formulated as finding an optimal sequence of matching vector pairs. Each vector pair is composed of an input vector that is matched to a model vector. Each pair has a pair matching score that reflects the matching quality between its vectors. The optimal sequence of vector pairs that pertain to an input vector sequence and a model vector sequence, has the highest sum of its pair matching scores compared to all other sequences of vector pairs from the same model and input sequences. The optimization algorithm gradually scans all the matching vector pairs, starting from pairs with the first input vector, and finds for each vector pair the optimal partial sequence of pairs that ends in that pair.[0064]
• The principle of Dynamic Programming (DP) states that any partial sequence of matching vector pairs with an optimal score includes only sub-partial sequences that are also optimal. It means that any pair that precedes the last pair of any optimal-partial sequence of vector pairs, is also an ending of an optimal-partial sequence of vector pairs. It implies that a DP-based algorithm has to gradually build partial sequences that each ends with a terminal vector pair by searching backwards and finding the preceding pair that contributes the most to the aggregate sum of matching scores that pertains to that terminal vector pair. Thus, the DP significantly reduces the computations required, since the algorithm has to find for each pair only one preceding pair that maximizes the aggregate score. The sequencing requirement described in paragraphs [0057][0058] further reduces the search space as elaborated in FIGS. 5A and 5B.[0065]
• The array A[0066]_j, constructed as described in FIG. 4, is used in this procedure. The basic idea in the approach is to process each cell of the array A_jand find the highest sum of pair-matching scores of a partial sequence that ends with matching pair (t_nk, m_rj), that pertains to that particular cell. A variable pointer1 holds the information of the current cell that is being processed. The pointer1 is initialized to the address of the first cell of array A_jin501 and is incremented in502 to process subsequent cells. Pointer1 always addresses the current cell whose partial vector sequence is being optimized. The procedure is repeated until all the cells have been considered in512. In order to find the sum of matching scores of a partial sequence that ends with matching pair (t_nk, m_rj), cells with input vector locations less than n and model vector location less than or equal to r have to be considered if they belong to the optimal partial sequence. During the process of consideration, cells generated due to matching vector pairs of t_(n−1)kare considered before pairs that pertain to t_(n−2)kand so on. This consideration process repeats itself until reaching the null vector t_0kand this is checked by511.
• While considering vector pairs with the input vector t[0067]_(n−1)k, only vector pairs that pertain to cells with R greater than 0 as the lower bound and R less than r as the higher bound are considered. But for pairs with subsequent input vectors the range could be even narrower with the lower bound tending towards r. This saves considerable computations as the number of cells considered in finding highest sum of matching scores of the partial sequence that ends with matching pair (t_nk, m_rj) is reduced considerably.
• Pointer[0068]2, addresses the preceding cells that are candidates to be included in the partial vector pair sequence that ends in the current cell. Pointer2 is initialized to the immediate preceding cell of current processed cell as shown in503. Pointer2 helps in navigating through the A_jas shown in505. In510, Pointer2 also helps in finding whether the checking has reached the first cell. In504, the algorithm checks for cells with required input vector location and allowable model vector locations. If the result of the checking in504 is positive, the candidate cells are also being checked if they qualify as a preceding matching vector pair in the partial sequence. This check is performed in525, by checking if the cell's current optimal sum of the matching scores of the partial sequence O_nkrjis smaller than the sum of matching vector pair score of the current cell W_nkrjwith the optimal sum O_nkrjof the cell whose address is given by pointer2. If the answer to the condition in525 is positive, the sum of the matching score of the partial sequence, stored in O_nkrjis incremented in526 by adding to the matching vector pair score of the current cell W_nkrj, the value of O_nkrjof the cell whose address is given by pointer2.
• The above process is performed for each A[0069]_jof array set {A_j} this takes care of the whole collection of models {M_j} that correspond to the set {A_j}. The cell among all the cells of {A_j} having the highest sum of matching score O_nkrjis selected as the highest sequence matching score and the input motion is recognized as most similar to the motion that corresponds to the winning model M_j.
• Preferred Embodiment of the Method of Speech Recognition[0070]
• Speech Recognition (SR) has been an active research objective for many years. At first, research concentrated on recognition of separate words in a limited vocabulary. More recently, research is focused on large vocabularies and continuous speech as well. Systems for speech recognition are already commercially available. However, even now those systems have high error rates and are limited to vocabularies of few thousand words in limited domains such as airline reservations.[0071]
• Currently, modern speech recognition systems rely on models that require hundreds of thousands parameters that are obtained by intensive statistical analysis of hundreds of hours of recorded speech. In contrast, in our proposed method of speech recognition, we aim at reducing significantly the required number of parameters that describe speech models. Our method is elaborated in this section. But, first we have to describe the prevalent SR methods and the basic concepts related to speech.[0072]
• The process of SR begins with sampling of the speech sound that is usually converted into electric signal by a microphone. This signal varies in its spectral contents relatively slowly due to physical limitations on the speed of motion of the speech articulators that generate the speech sounds, i.e. the glottis, lips and tongue. Therefore, it is possible to sample the speech signal at relatively slow frequency typically at about 100 Hz and to produce from each sample an Acoustic Feature Vector (AFV) that represents the local spectral contents of each sample. The original speech signal is now represented by a series of AFVs. Each AFV is actually a vector with approximately 40 dimensions, where each dimension represents the amplitude of one of the spectral components of the speech sampled at that point in time.[0073]
• The basic SR problem is to recognize the set of utterances (spoken words) that could have generated a given sequence of AFVs. To simplify the solution of such a problem one may choose to limit the speaker's vocabulary. Reducing the size of the lexicon that lists the vocabulary, could reduce the complexity of the SR algorithm tremendously. Each utterance can be divided into phonemes, which are individual speech sounds used to pronounce each word. The phonemes could be regarded as the basic units of speech. The AFVs, which are much smaller in their time span (few milliseconds), actually are small pieces of the phonemes that generated them.[0074]
• Practical lexicons typically contain tens of thousands of words. In practice, not only the words uttered by the speaker are a priori unknown, but also the number of words in each sequence. However, language constraints strongly influence the probability of various word sequences. For example, the sequence “we are a family” is much more plausible than “are family a we”. In addition, more clues for word recognition are provided by the sequence of phonemes, that are in turn composed of sequences of AFVs. In summary, the complete representation of speech can be depicted as a four layer pyramid. The lower layer is composed of AFVs that are the elementary components of phonemes, the second layer is composed of phonemes that are the basic parts of words that are represented by the third layer. Evidently, the top layer represents sentences composed of sequences of words.[0075]
• In order to recognize words from a sequence of AFVs, the SR system has to find the word with the highest probability that has been uttered, given a particular sequence of AFVs. Most of the SR systems rely on the maximum likelihood principle, i.e. maximizing the conditional probability P(y|w) of an AFV sequence y given a particular word w. Multiplying P(y|w) by P(w) is equal according to Bayes law to P(w|y)P(y). Assuming that P(y) is the same for all y's, maximizing P(w|y) is equivalent to maximizing P(y|w)P(w). The recognized word is the word w with the highest a-posteriori probability P(w|y), but this probability is very difficult to estimate. On the other hand, both P(y|w) and P(w) can be estimated more easily. P(w) can be derived from a model that describes the relative frequencies of the various vocabulary words based on a statistical model of the language. Determining the other term i.e. P(y|w) is much more difficult due to the stochastic and variable nature of speech, which causes the same word to be plausibly represented by a large variety of APV sequences. In addition to the natural stochastic variation and temporal variations of the sequences, there are different pronunciations of the same word with different phonemes or co-articulation that depends on neighboring words in the sentences or even on neighboring phonemes. To compensate for the temporal variations of the speech signal, the APVs used in SR may include also first and second temporal derivatives of their spectral components.[0076]
• Most contemporary SR systems employ the Hidden Markov Model (HMM) to find the most probable word w given an APV sequence y, i.e. to find a w that maximizes [P(w|y)]. The HMM models the stochastic nature of speech and operates as a probabilistic finite state machine. This machine can be described as a directed graph. The graph nodes represent states and the graph's directed arcs represent feasible transitions from one state to the next. The arcs can also start and end at the same state (self loop). Thus, the topology of the graph determines which transitions are valid and which are not. If two nodes are not connected, the machine can not transit between these nodes. Also, each of the transitions may have different probabilities. So, each arc has an attached weight that signifies the probability of its transition. At each time frame the machine makes a probabilistic transition. Each transition generates one APV. To emit a full phoneme, one has to traverse the graph from the initial state to the final state. Due to the probabilistic structure of the HMM the same phoneme could have a different sequence of APVs each time. In practice, these variations actually represent different pronunciations of the same phoneme while the sequences of APVs represent their temporal-spectral contents.[0077]
• If the HMM graph represents a phoneme, the different ways to pronounce it are modeled by the various ways one can traverse this graph from the first node to the last. The transition probabilities attached to the arcs govern the statistical distribution of these pronunciations. The duration of the phoneme is equal to the number of transitions (and also time frames) required to traverse the graph. The transition probabilities that govern the behavior of the HMM for each phoneme, has to be estimated from a very large set of spoken words, which are analyzed by extracting their APVs from the speech signal and then applying special maximum likelihood (ML) algorithms. Actually, the situation is even more complicated since the probabilities are modeled as probability densities and the ML estimates their parameters. For example, if the probability densities are modeled as a mixture of Gaussians, the required parameter set includes the means and the standard deviations of these Gaussians.[0078]
• As already mentioned, the estimation process of the parameters of the transition probability densities is very complex. The process is called training of the particular HMM that is representing a specific utterance. Such HMM is composed of concatenated HMMs that each represents a phoneme in the utterance that may include more than one word. The estimation process necessitates the use of the Expectation Maximization (EM) algorithm. The situation is further complicated by effects such as co-articulation. In such a case, the sequence of APVs that describe a phoneme in an utterance may vary depending on the neighboring phonemes i.e. the context. This requires to find a huge amount of densities to represent all the possibilities. For example, if a language has 50 different phonemes, and one considers as context only one phoneme preceding and one trailing phoneme as a context that affects the co-articulation, the number of possible variations rises to 125,000. To manage such an enormous complexity, the densities have to be clustered and represented only by the few probability densities that are at the centers of these clusters.[0079]
• After training the probabilities, given a sequence of APVs that are derived from an utterance, one has to find the sequence of words that were uttered. This is the main goal of the SR algorithm. The major use of HMMs in SR is to compute the likelihood that a given sequence of AFVs y is produced by the HMM of a particular phoneme w (or a word). To infer which phoneme generated a given AFV sequence, one has to compare the joint probabilities P(y|w)P(w) i.e. all the likelihood's that different w's generated y. The phoneme (or word) w that has the highest likelihood is the one that best matches the given AFV sequence y. The Viterbi algorithm, which is based on dynamic programming is used to find the highest likelihood for each phoneme. After finding the most probable phoneme foe each AFV sequence, it is necessary to find which word has the highest likelihood to represent the phoneme sequence found. Again, this procedure is quite lengthy since one has to compute the likelihood's of all the w HMMs in the vocabulary for a given sequence y. Such a procedure is serial because it requires to compute all the possible w HMMs in series and then to compare them. In comparison, our method is much more efficient since it is parallel, i.e. it considers and compares all the possible w's at once.[0080]
• In our approach, which is called Recognition by Indexing and Sequencing (RISq), one can vote on phonemes or even directly on words. An hierarchical voting structure, where initially one uses the AFVs to vote for phonemes and then uses the identified phonemes to vote for the most probable words, is also feasible.[0081]
• Assuming that our RISq voting is performed with AFVs, it is necessary to vote for each AFV only once. After the first AFV is voted, the next AFV in the sequence y is also voted (but on a smaller word/phoneme set) and so on. Each vote yields all the possible words (or phonemes) that could include the voting AFV in their signal. The first AFV votes for all the possible words (or phonemes) in the vocabulary. The second AFV votes only on words (or phonemes) that include the second AFV after the first AFV. The third AFV votes on an even more reduced set of words (or phonemes) that only include the third AFV after the second AFV, and so on. Each AFV votes on a reduced set of words/phonemes than the previous AFV. This is our sequencing approach that conserves the sequence but still allows missing AFVs (unuttered) in the sequence y, as frequently occurs in practice. This voting procedure has to be performed only once. At the end of the voting, the word/phoneme w with the highest voting score is identified as the most probable word/phoneme uttered by the sequence y.[0082]
• This process is more efficient than the HMM approach. Since AFVs almost never match exactly the indexes in the table, the voting score can be adjusted in an inversely proportional relation to the multidimensional distance between the voting AFV and the several nearest indexes in the table. Such an inversely proportional relation can be defined as various functions of the distance, such as the Gaussian function of the Mahalanobis distance. These functions depend on the joint probability density of the vector components (dimensions) of the voting AFV. A multidimensional Gaussian function is the most plausible joint density for the preferred embodiment, but other functions could be used as well.[0083]
• The training of our RISq actually requires only to construct the voting table. This table is arranged as a set of bins, where each bin is indexed (addressed) by another AFV and each bin contains a list of all the words/phonemes that include in their sequence y the bin's AFV. It is suggested to employ a Self Organizing Mapping (SOM) algorithm to find the best AFVs to represent the vocabulary. Even if the construction process of the voting table is quite tedious and lengthy, it has to be performed only once. Once the voting table is finished, the voting process could be quite fast.[0084]
• As mentioned above, we could vote hierarchically first on phonemes from AFVs and then on words from phonemes. Such a process could significantly simplify the complexity of the voting tables relatively to tables required for direct voting for words from AFVs. We shall test both methods empirically. But, the principles of RISq method for both approaches remain the same.[0085]

Claims (45)

What is claimed is:

1. A method for recognizing an input human motion as most similar to a model human motion, which is a member of a stored collection of model human motions, comprising the steps of:

a. Creating a collection of model human motions by measuring and recording the model trajectories of body parts that pertain to human performances of such collection of model human motions;

b. sampling at a model-rate of sampling, one recorded model trajectories of body parts;

c. repeating step b each time with another recorded model trajectories of body parts, until all the recorded model trajectories of body parts included in the collection of model human motions, have been sampled;

d. representing each sample of the recorded model trajectories of body parts by a model vector m_rj; wherein each component of such model vector is derived from a sample of a recorded model trajectory of one body part;

e. representing each member of the collection of model human motions by a model vectors sequence M_j=(m_1j. . . m_rj. . . m_qj); wherein the subscript q denotes the total number of model vectors in the model vectors sequence M_j; wherein the subscript r denotes the location of the model vector m_rjwithin the model vectors sequence M_j; wherein the subscript j denotes the serial number of the model vectors sequence M_jwithin the collection of model vectors sequences {M_j} that represents the corresponding collection of model human motions;

f. storing in a hash table, the entire collection of model vectors sequences {M_j} by storing each of the model vectors m_rjthat belongs to the collection of model vectors sequences {M_j}, in a hash table bin whose address is the nearest to the model vector m_rj;

g. Acquiring an input human motion by measuring and recording the input trajectories of body parts that pertain to a human performance of such input human motion;

h. sampling at an input-rate of sampling, the recorded input trajectories of body parts; wherein the input-rate of sampling is set to be substantially different from the model-rate of sampling;

i. representing each sample of the recorded input trajectories of body parts by an input vector t_nk; wherein each component of such an input vector is derived from a sample of a recorded input trajectory of the same body part that pertains to the corresponding component of the model vector m_rj;

j. representing the sampled input human motion by an input vectors sequence T_k=(t_1k. . . t_nk. . . t_pk); wherein the subscript p denotes the total number of input vectors in the input vectors sequence; wherein the subscript n denotes the location of the input vector t_nkwithin the input vectors sequence T_k; wherein the subscript k denotes the serial number of the input vectors sequence;

k. employing an optimal matching algorithm to find the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j};

l. repeating step k until all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} are found;

m. comparing all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} and finding one model vectors sequence M_othat has the highest optimal matching score; recognizing the input human motion as most similar to the model human motion that pertains to the model vectors sequence M_owith the highest optimal matching score;

2. The method ofclaim 1 wherein employing the optimal matching algorithm for finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. indexing into a hash table, each input vector t_nkthat belongs to the input vectors sequence T_k, and retrieving all the model vectors m_rjthat are near to each input vector t_nk; using the model vectors retrieved to construct all the vector pairs (t_nk, m_rj) composed of vectors that are near to one another;

b. computing for each vector pair (t_nk, m_rj) retrieved in step a, a pair matching score that reflects the vector pair's matching quality;

c. removing from all the vector pairs retrieved all the vector pairs that have a pair matching score equal or below a predetermined threshold and retaining the remaining vector pairs and naming these remaining vector pairs as matching vector pairs;

d. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence; wherein the valid pair sequence is constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, m_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions: (I) c is not equal to n (II) if c>n, then r≦e; (III) if c<n then e≦r;

e. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

3. The method ofclaim 2 wherein the recorded trajectories of human body parts are equivalent to recorded poses of human body parts; wherein both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from samples of recorded poses of human body parts and from temporal derivatives of samples of recorded poses of human body parts.

4. The method ofclaim 3 wherein the samples of recorded poses of human body parts and the temporal derivatives of samples of recorded poses of human body parts, correspond to samples of recorded angular poses of body parts and to temporal derivatives of samples of recorded angular poses of human body parts.

5. The method ofclaim 4 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

6. The method ofclaim 5 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

7. The method ofclaim 4 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and each model vectors sequence M_j, which is a member of the collection of model vectors sequences {M_j} is acquired by slow-sparse model-rate of sampling.

8. The method ofclaim 7 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

9. The method ofclaim 1 wherein employing the optimal matching algorithm for finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. indexing into a hash table, each input vector t_nkthat belongs to the input vectors sequence T_nk, and retrieving all the model vectors m_rjthat are near to each input vector t_nk; using the model vectors retrieved to construct all the vector pairs (t_nk, m_rj) composed of vectors that are near to one another;

b. computing for each vector pair (t_nk, m_rj) retrieved in step a, a pair matching score that reflects the vector pair's matching quality;

c. removing from all the vector pairs retrieved all the vector pairs that have a pair matching score equal or below a predetermined threshold and retaining the remaining vector pairs and naming these remaining vector pairs as matching vector pairs;

d. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence; wherein the valid pair sequence is constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, M_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions: (I) r is not equal to e; (II) if e>r, then n≦c; (III) if e<r then c≦n;

e. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

10. The method ofclaim 9 wherein the recorded trajectories of human body parts are equivalent to recorded poses of human body parts; wherein both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from samples of recorded poses of human body parts and from temporal derivatives of samples of recorded poses of human body parts.

11. The method ofclaim 10 wherein the samples of recorded poses of human body parts and the temporal derivatives of samples of recorded poses of human body parts, correspond to samples of recorded angular poses of body parts and to temporal derivatives of samples of recorded angular poses of human body parts.

12. The method ofclaim 11 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

13. The method ofclaim 12 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

14. The method ofclaim 11 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and each model vectors sequence M_j, which is a member of the collection of model vectors sequences {M_j} is acquired by slow-sparse model-rate of sampling.

15. The method ofclaim 14 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

16. A method for recognizing an input vectors sequence T_kas most similar to a model vectors sequence M_j, which is a member of a stored collection of model vectors sequences {M_j}, comprising the steps of:

a. Creating a collection of model signal sets by measuring and recording all the model signal sets of a collection of model signal sets;

b. sampling at a model-rate of sampling, one model signal set;

c. repeating step b each time with another model signal set, until all the recorded model signal sets included in the collection of model signal sets, have been sampled;

d. representing each sample of the model signal set by a model sample vector w_rj; wherein each component of such model sample vector is derived from one sample of one recorded model signal that belongs to the model signal set;

e. transforming each model sample vector w_rjinto a model vector m_rjusing a pre-specified transformation;

f. representing each member of the collection of model vectors sequences by a model vectors sequence M_j=(m_1j. . . m_rj. . . m_qj); wherein the subscript q denotes the total number of model vectors in the model vectors sequence M_j; wherein the subscript r denotes the location of the model vector m_rjwithin the model vectors sequence M_j; wherein the subscript j denotes the serial number of the model vectors sequence M_jwithin the collection of model vectors sequences {M_j};

g. storing in a hash table, the entire collection of model vectors sequences {M_j} by storing each of the model vectors m_rjthat belongs to the collection of model vectors sequences {M_j}, in a hash table bin whose address is the nearest to the model vector m_rj;

h. Acquiring an input signal set by measuring and recording an input signal set; wherein each input signal of the input signal set corresponds to a model signal of the model signal set;

i. sampling at an input-rate of sampling, the recorded input signal set; wherein the input-rate of sampling is adjusted to be substantially different from the model-rate of sampling;

j. representing each sample of the recorded input signal set by an input sample vector v_nk; wherein each component of such an input sample vector is derived from one sample of a recorded input signal that belongs to the input signal set;

k. transforming each input sample vector v_nkinto an input vector t_rjusing a pre-specified transformation;

l. representing the sampled input signal set by an input vectors sequence T_k=(t_1k. . . t_nk. . . t_pk); wherein the subscript p denotes the total number of input vectors in the input vectors sequence; wherein the subscript n denotes the location of the input vector t_nkwithin the input vectors sequence T_k; wherein the subscript k denotes the serial number of the input vectors sequence;

m. employing an optimal matching algorithm to find the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j};

n. repeating step k until all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} are found;

o. comparing all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} and finding one model vectors sequence M_othat has the highest optimal matching score; recognizing the input signal set as most similar to the model signal set that pertains to the model vectors sequence M_owith the highest optimal matching score;

17. The method ofclaim 16 wherein employing the optimal matching algorithm for finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. indexing into a hash table, each input vector t_nkthat belongs to the input vectors sequence T_k, and retrieving all the model vectors m_rjthat are near to each input vector t_nkusing the model vectors retrieved to construct all the vector pairs (t_nk, m_rj) composed of vectors that are near to one another;

b. computing for each vector pair (t_nk, m_rj) retrieved in step a, a pair matching score that reflects the vector pair's matching quality;

c. removing from all the vector pairs retrieved all the vector pairs that have a pair matching score equal or below a predetermined threshold and retaining the remaining vector pairs and naming these remaining vector pairs as matching vector pairs;

d. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence; wherein the valid pair sequence is constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, M_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions:

(I) c is not equal to n (II) if c>n, then r≦e; (III) if c<n then e≦r

e. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

18. The method ofclaim 17 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

19. The method ofclaim 18 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

20. The method ofclaim 19 wherein both the input vectors sequence T_kand the collection of model vectors sequences {M_j} represent speech utterances and both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from values of sampled speech signals, discrete transforms of sampled speech signals and their temporal derivatives.

21. The method ofclaim 17 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and acquiring by a slow-sparse model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

22. The method ofclaim 21 wherein the highest sequence matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

23. The method ofclaim 22 wherein both the input vectors sequence T_kand the collection of model vectors sequences {M_j} represent speech utterances and both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from values of sampled speech signals, discrete transforms of sampled speech signals and their temporal derivatives.

24. The method ofclaim 16 wherein employing the optimal matching algorithm for finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. indexing into a hash table, each input vector t_nkthat belongs to the input vectors sequence T_k, and retrieving all the model vectors m_rjthat are near to each input vector t_nk; using the model vectors retrieved to construct all the vector pairs (t_nk, m_rj) composed of vectors that are near to one another;

b. computing for each vector pair (t_nk, m_rj) retrieved in step a, a pair matching score that reflects the vector pair's matching quality;

c. removing from all the vector pairs retrieved all the vector pairs that have a pair matching score equal or below a predetermined threshold and retaining the remaining vector pairs and naming these remaining vector pairs as matching vector pairs;

d. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence; wherein the valid pair sequence is constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, m_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions:

(I) r is not equal to e; (II) if e>r, then n≦c; (III) if e<r then c≦n;

e. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

25. The method ofclaim 24 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

26. The method ofclaim 25 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

27. The method ofclaim 26 wherein both the input vectors sequence T_kand the collection of model vectors sequences {M_j} represent speech utterances and both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from values of sampled speech signals, discrete transforms of sampled speech signals and their temporal derivatives.

28. The method ofclaim 24 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and acquiring by a slow-sparse model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

29. The method ofclaim 28 wherein the highest sequence matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

30. The method ofclaim 29 wherein both the input vectors sequence T_kand the collection of model vectors sequences {M_j} represent speech utterances and both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from values of sampled speech signals, discrete transforms of sampled speech signals and their temporal derivatives.

31. A method for recognizing an input human motion as most similar to a model human motion, which is a member of a stored collection of model human motions, comprising the steps of:

a. Creating a collection of model human motions by measuring and recording the model trajectories of body parts that pertain to human performances of such collection of model human motions;

b. sampling at a model-rate of sampling, one recorded model trajectories of body parts;

c. repeating step b each time with another recorded model trajectories of body parts, until all the recorded model trajectories of body parts included in the collection of model human motions, have been sampled;

d. representing each sample of the recorded model trajectories of body parts by a model vector m_rj; wherein each component of such model vector is derived from a sample of a recorded model trajectory of one body part;

e. representing each member of the collection of model human motions by a model vectors sequence M_j=(m_1j. . . m_rj. . . m_qj); wherein the subscript q denotes the total number of model vectors in the model vectors sequence M_j; wherein the subscript r denotes the location of the model vector m_rjwithin the model vectors sequence M_j; wherein the subscript j denotes the serial number of the model vectors sequence M_jwithin the collection of model vectors sequences {M_j} that represents the corresponding collection of model human motions;

f. Acquiring an input human motion by measuring and recording the input trajectories of body parts that pertain to a human performance of such input human motion;

g. sampling at an input-rate of sampling, the recorded input trajectories of body parts; wherein the input-rate of sampling is set to be substantially different from the model-rate of sampling;

h. representing each sample of the recorded input trajectories of body parts by an input vector t_nk; wherein each component of such an input vector is derived from a sample of a recorded input trajectory of the same body part that pertains to the corresponding component of the model vector m_rj;

i. representing the sampled input human motion by an input vectors sequence T_k=(t_1k. . . t_nk. . . t_pk); wherein the subscript p denotes the total number of input vectors in the input vectors sequence; wherein the subscript n denotes the location of the input vector t_nkwithin the input vectors sequence T_k; wherein the subscript k denotes the serial number of the input vectors sequence;

j. dividing the full set of human body parts into several groups of human body parts;

k. in complete correspondence to the division into groups of human body parts, also dividing each of the input vectors and each of the model vectors into input sub-vectors and corresponding model sub-vectors; wherein each sub-vector pertains to another group of human body parts;

l. storing each division of model sub-vectors, which corresponds to the same group of body parts, in a separate hash sub-table; wherein the hash sub-table has addressing sub-vectors that correspond in their dimensions to the dimensions of the model sub-vectors stored; wherein each model sub-vector is stored in a hash sub-table bin whose address sub-vector is the nearest to the model sub-vector; wherein the bins have address sub-vectors that are determined by a process of vector quantization;

m. indexing each of the input vectors t_nkwithin the input vectors sequence T_kby using each of its input sub-vectors to index separately into the corresponding hash sub-table;

n. for each input vector t_nkwithin the input vectors sequence T_k, retrieving the corresponding model sub-vectors extracted by separately indexing with input sub-vectors that are parts of the same input vector t_nkand merging them back into model vectors m_rj; creating from each retrieved and merged model vector m_rja vector pair (t_nk, m_rj);

o. computing for each vector pair (t_nk, m_rj) created in step n, a pair matching score that reflects the vector pair's matching quality;

p. removing from all the vector pairs created in step n, all the vector pairs that have a pair matching score equal or below a predetermined threshold and retaining the remaining vector pairs and naming these vector pairs as matching vector pairs;

q. employing an optimal matching algorithm to find the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_jwhich is one of the members of the collection of model vectors sequences {M_j};

r. repeating step q until all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} are found;

s. comparing all the optimal matching scores that pertain to all the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j} and finding one model vectors sequence M_nthat has the highest optimal matching score; recognizing the input human motion as most similar to the model human motion that pertains to the model vectors sequence M_nwith the highest optimal matching score;

32. The method ofclaim 31 wherein finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. constructing from the matching vector pairs, all the permutations of valid pair sequences; wherein all the valid pair sequences are constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, m_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions: (I) c is not equal to n (II) if c>n, then r≦e; (III) if c<n then e≦r;

b. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence;

c. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

33. The method ofclaim 32 wherein the recorded trajectories of human body parts are equivalent to recorded poses of human body parts; wherein both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from samples of recorded poses of human body parts and from temporal derivatives of samples of recorded poses of human body parts.

34. The method ofclaim 33 wherein the samples of recorded poses of human body parts and the temporal derivatives of samples of recorded poses of human body parts, correspond to samples of recorded angular poses of body parts and to temporal derivatives of samples of recorded angular poses of human body parts.

35. The method ofclaim 34 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

36. The method ofclaim 35 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

37. The method ofclaim 34 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and acquiring by a slow-sparse model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

38. The method ofclaim 37 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

39. The method ofclaim 31 wherein finding the optimal matching score between the input vectors sequence T_kand one model vectors sequence M_j, which is one of the members of the collection of model vectors sequences {M_j}, comprising of the steps:

a. constructing from the matching vector pairs, all the permutations of valid pair sequences; wherein all the valid pair sequences are constructed only from sequences of matching vector pairs that have mutual sequential relation; wherein the mutual sequential relation between any two matching vector pairs . . . (t_nk, m_rj) . . . (t_ck, m_ej) . . . included in the valid pair sequence has to fulfill the following 3 conditions: (I) r is not equal to e; (II) if e>r, then n≦c; (III) if e<r then c≦n;

b. employing an optimization algorithm to find the valid pair sequence with the highest sequence matching score; wherein a sequence matching score of a valid pair sequence is defined as the sum of all the pair matching scores that pertain to the matching vector pairs that are components of the valid pair sequence;

c. naming the valid pair sequence with the highest sequence matching score as the optimal pair sequence with the optimal matching score;

40. The method ofclaim 39 wherein the recorded trajectories of human body parts are equivalent to recorded poses of human body parts; wherein both the collection of model vectors sequences {M_j} and the input vectors sequence T_kare derived from samples of recorded poses of human body parts and from temporal derivatives of samples of recorded poses of human body parts.

41. The method ofclaim 40 wherein the samples of recorded poses of human body parts and the temporal derivatives of samples of recorded poses of human body parts, correspond to samples of recorded angular poses of body parts and to temporal derivatives of samples of recorded angular poses of human body parts.

42. The method ofclaim 41 wherein the input vectors sequence T_kis acquired by a sparse-slow input-rate of sampling and acquiring by a fast model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

43. The method ofclaim 42 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

44. The method ofclaim 41 wherein the input vectors sequence T_kis acquired by a fast input-rate of sampling and acquiring by a slow-sparse model-rate of sampling each of the model vectors sequences M_j, which are members of the collection of model vectors sequences {M_j}.

45. The method ofclaim 44 wherein the optimal matching scores are obtained by a special optimization algorithm that employs principles of dynamic programming.

Images